Cosine, an artificial intelligence startup that’s focused on mimicking human reasoning, said today it has closed on a $2.5 million seed funding round to finance the development of its cutting-edge AI software developer.
Today’s round was led by U.S. based venture capital firms Uphonest and SOMA Capital, and saw participation from Lakestar, Focal and others.
The startup, officially known as Buildt AI Inc., also revealed that its most advanced model to date has achieved what it claims is the highest-ever score achieved by any AI model so far on SWE-Bench, an industry standard benchmark for evaluating AI software engineering skills.
According to Cosine, its AI-powered software developer recently scored 30% on the SWE-Bench test. That might not sound like a lot, but it represents a 56% improvement over the next best score of just 19%, achieved by The San Francisco AI Factory Inc., and an incredible 2,196% improvement on OpenAI’s GPT-4, which scored just 1.31% on the test.
The SWE-Bench benchmark challenges AI models to perform a series of software architecture tasks, such as debugging problems and implementing new features in existing codebases. As they do this, it assesses their ability to understand, generate and modify existing complex code.
The startup introduced its most advanced AI developer as Genie, and said it’s the equivalent of a “very good human developer.” It’s said to possess the ability to solve bugs, build new software features, refactor code and perform many other coding-related tasks, either fully autonomously or in collaboration with human developers.
Cosine says its secret sauce is that it has fine-tuned Genie to “emulate human reasoning,” and that this approach also enabled it to surpass better-known AI developer models such as $20 billion-valued Cognition AI Inc.’s Devin and Amazon Web Services Inc.’s Amazon Q Developer, which both scored less than 20% on the SWE-Bench test.
Cosine was co-founded by its Chief Executive Alistair Pullen, Chief Operating Officer Yang Li and Chief Information Officer Sam Stenner, who said they first realized the potential of large language models to imitate human software developers in early 2022. Stenner explained that they did this by codifying human reasoning, before using that code to train Genie’s underlying large language model.
The result is that Genie is “uncannily human” in its approach to reasoning, setting a new standard as the “world’s most human-like autonomous AI software developer.” The co-founders added that it’s designed for companies that want to expand their developer teams without going through the hassles and friction of hiring a human.
Pullen said the company’s breakthroughs in human reasoning mean it has been able to build AI models that can “operate far beyond the narrow range of tasks and tightly restricted prompts currently available to teams developing software.”
Stenner stressed that Cosine isn’t looking to replace human developers entirely, but rather augmented them with human-like assistants that they can collaborate with on any kind of coding task.
“We are focused on creating a colleague, not a co-pilot,” he insisted. “We knew the potential for what we had built and worked with OpenAI to fine tune their largest context window LLMs. We’re confident we now have the capabilities to consistently beat our own top score.”